vadiml1024/signal-analysis

signal-analysis is a Claude Skill, a set of instructions you can load into Claude so it follows a specific method when you ask it to look at a claim, headline, filing, announcement, or a pile of unstructured chatter. Instead of answering off the cuff, Claude works through eight structured steps to figure out what the information actually means. The skill is built around eleven working rules borrowed from staff analyst practice. These include checking who said something and why, separating fact from assessment from conclusion, tracing motives behind a message, weighing capabilities over stated intentions, watching for changes in language over time, comparing trends instead of single snapshots, treating silence, such as a metric that stops being published or a denial that never comes, as meaningful, and resisting the urge to fall in love with one's own hypothesis. It also stresses that repeating one source through multiple outlets counts as a single signal, not several independent ones. When triggered, the skill walks through eight steps: taking inventory of the input and discarding noise, breaking the material into facts versus assessments versus conclusions, mapping who benefits from the claim being believed, checking whether stated capabilities match the claim, sweeping for anomalies like language drift or numbers that do not add up, mapping how many genuinely independent sources corroborate the claim, building a rival hypothesis to argue against the first one, and producing a calibrated conclusion with a confidence level and a list of things to watch next. The output follows a fixed template that includes arguments against its own hypothesis, what to watch, and gaps in the analysis. It installs either through Claude.ai or Claude Desktop by uploading a packaged skill file, or through Claude Code by cloning the repository into a skills folder. The project notes that any material being analyzed is treated strictly as data, not as instructions, so directives buried inside a document under review are logged rather than obeyed. The README states the rules are adapted from a Russian language essay by a former general staff officer, with the workflow and supporting reference material built out from that original set of ideas. The project is released under the MIT license.